@pome-sh/cli 0.23.26 → 0.23.27

package/dist/build-info.json

@@ -1,6 +1,6 @@
 {
   "package": "pome-sh",
-  "version": "0.23.26",
-  "git_sha": "0e181da493759acb3bfe57f63a3c623389e7e394",
-  "build_time": "2026-08-12T19:09:17.801Z"
+  "version": "0.23.27",
+  "git_sha": "87e466ebca9fffe55fab831bbb52efaf68dce110",
+  "build_time": "2026-08-12T21:42:53.518Z"
 }

package/dist/{checks-UINTOJ2L.js → checks-WHR4ALIJ.js}

@@ -183,7 +183,7 @@ async function checkTwinReachable(_configDir) {
   });
   let harness;
   try {
-    const { bootTwin } = await import('./twinHarness-WVKOQ7UP.js');
+    const { bootTwin } = await import('./twinHarness-H73PEFYS.js');
     harness = await bootTwin({
       twin: "github",
       seedState: void 0,

package/dist/{chunk-DINCNFAQ.js → chunk-4LODFVKV.js}

@@ -1,5 +1,5 @@
 import { readManifest, normalizeManifestTwins } from './chunk-ABM3CMQB.js';
-import { createHostedClient, perTwinReturnedByCloud, parseTaskFile, runAgentCommand, writeRunArtifactsCore, toTwinHttpEvent, redactJsonl, uploadRunBlobs, scoreFromFinalizeResponse, scoreStatus, evaluationCounts } from './chunk-VBLY7LYM.js';
+import { createHostedClient, perTwinReturnedByCloud, parseTaskFile, runAgentCommand, writeRunArtifactsCore, toTwinHttpEvent, redactJsonl, uploadRunBlobs, scoreFromFinalizeResponse, scoreStatus, evaluationCounts } from './chunk-XKBL7JKK.js';
 import { MOUNTED_TWINS, HostedAuthError, HostedDiscardRefusedError, HostedQuotaError, HostedOrchError, HostedTrialError, agentResponseSchema } from './chunk-X66JOOO7.js';
 import { redactSecrets, redactEvent } from './chunk-SG6ZTIMT.js';
 import { existsSync } from 'node:fs';

package/dist/{chunk-4VVQMRWP.js → chunk-CC2N4JGQ.js}

@@ -84,9 +84,9 @@ var TWIN_REGISTRY = {
     envName: "STRIPE",
     defaultPort: 3333,
     version: package_default3.version,
-    defaultSeed: async () => (await import('./src-6JNLHNYE.js')).defaultSeed(),
+    defaultSeed: async () => (await import('./src-HRW7OZBS.js')).defaultSeed(),
     async boot({ seedState, runId, recorder, twinBaseUrl }) {
-      const stripeTwin = await import('./src-6JNLHNYE.js');
+      const stripeTwin = await import('./src-HRW7OZBS.js');
       const { createApp } = await import('./server-KOR7UHEP.js');
       const {
         applySeed: applyStripeSeed,

package/dist/{chunk-DG2VDEQE.js → chunk-FGNIWUEU.js}

@@ -1,4 +1,4 @@
-import { TWIN_NAMES, isTwinName, TWIN_REGISTRY } from './chunk-4VVQMRWP.js';
+import { TWIN_NAMES, isTwinName, TWIN_REGISTRY } from './chunk-CC2N4JGQ.js';
 import { createFileBackedRecorderStore, createRecorderStore } from './chunk-OW7VVW6X.js';
 
 // src/recorder/recorder.ts

package/dist/{chunk-ZUQYEPY4.js → chunk-QIVE6VJE.js}

@@ -1,7 +1,7 @@
 import { getAvailablePort } from './chunk-XDU6TD4O.js';
 import { buildEgressAllowlist, readBlockedEgress } from './chunk-CBFKZZBR.js';
-import { parseTaskFile, seedStateForTwin, runAgentCommand, writeRunArtifactsCore } from './chunk-VBLY7LYM.js';
-import { createRecorder, bootTwin } from './chunk-DG2VDEQE.js';
+import { parseTaskFile, seedStateForTwin, runAgentCommand, writeRunArtifactsCore } from './chunk-XKBL7JKK.js';
+import { createRecorder, bootTwin } from './chunk-FGNIWUEU.js';
 import { eventSchema } from './chunk-VBATFCWR.js';
 import { redactSecrets, redactEvent } from './chunk-SG6ZTIMT.js';
 import { serve } from '@hono/node-server';

package/dist/{chunk-VBLY7LYM.js → chunk-XKBL7JKK.js}

@@ -2,7 +2,7 @@ import { criterionSchema, finalizeResponseSchema, HostedDiscardRefusedError, Hos
 import { seedSchema, parseSeed, defaultSeedState as defaultSeedState$2 } from './chunk-KIUIKAVU.js';
 import { gmailSeedSchema, defaultSeedState } from './chunk-NJ246QPJ.js';
 import { linearSeedSchema, defaultSeedState as defaultSeedState$1 } from './chunk-ZKID2HS3.js';
-import { isTwinName, TWIN_REGISTRY } from './chunk-4VVQMRWP.js';
+import { isTwinName, TWIN_REGISTRY } from './chunk-CC2N4JGQ.js';
 import { toTwinHttpEventRow } from './chunk-OW7VVW6X.js';
 import { redactEvent, redactSecrets } from './chunk-SG6ZTIMT.js';
 import { mkdir, appendFile, writeFile, readFile } from 'node:fs/promises';

package/dist/{runDemo-5Z7PIYTZ.js → runDemo-HFISTHIE.js}

@@ -1,16 +1,16 @@
 import { newGroupId, reassuranceBox, twinReadyLine, trialsHeaderLine, trialLine, summaryLines, evaluatingLine, criterionPhrase } from './chunk-RGZBC7NF.js';
 import { DemoCapacityError, capacityLabel, parseCapacityMarker, capacityKindFrom } from './chunk-ZX4WNSZ5.js';
-import { runTask, demoTaskPath, DEMO_TASK_NAME, DEMO_REPO } from './chunk-ZUQYEPY4.js';
+import { runTask, demoTaskPath, DEMO_TASK_NAME, DEMO_REPO } from './chunk-QIVE6VJE.js';
 import { getAvailablePort } from './chunk-XDU6TD4O.js';
 import './chunk-CBFKZZBR.js';
-import { createHostedClient, parseTaskFile, uploadRunBlobs, scoreFromFinalizeResponse, scoreStatus, outcomeOf } from './chunk-VBLY7LYM.js';
+import { createHostedClient, parseTaskFile, uploadRunBlobs, scoreFromFinalizeResponse, scoreStatus, outcomeOf } from './chunk-XKBL7JKK.js';
 import './chunk-NW7HGA2K.js';
 import { HostedQuotaError, HostedOrchError } from './chunk-X66JOOO7.js';
 import './chunk-KIUIKAVU.js';
 import './chunk-NJ246QPJ.js';
 import './chunk-ZKID2HS3.js';
-import { bootTwin } from './chunk-DG2VDEQE.js';
-import './chunk-4VVQMRWP.js';
+import { bootTwin } from './chunk-FGNIWUEU.js';
+import './chunk-CC2N4JGQ.js';
 import './chunk-OW7VVW6X.js';
 import './chunk-VBATFCWR.js';
 import './chunk-SG6ZTIMT.js';

package/dist/{runTrialGroup-CL4EJZD2.js → runTrialGroup-LBNWOZPA.js}

@@ -1,13 +1,13 @@
 import { newGroupId, criterionPhrase } from './chunk-RGZBC7NF.js';
-import { runTaskHosted, resolveRunAgentIdentity } from './chunk-DINCNFAQ.js';
+import { runTaskHosted, resolveRunAgentIdentity } from './chunk-4LODFVKV.js';
 import './chunk-ABM3CMQB.js';
-import { createHostedClient, parseTaskFile, outcomeOf } from './chunk-VBLY7LYM.js';
+import { createHostedClient, parseTaskFile, outcomeOf } from './chunk-XKBL7JKK.js';
 import './chunk-NW7HGA2K.js';
 import { HostedQuotaError, HostedTrialError } from './chunk-X66JOOO7.js';
 import './chunk-KIUIKAVU.js';
 import './chunk-NJ246QPJ.js';
 import './chunk-ZKID2HS3.js';
-import './chunk-4VVQMRWP.js';
+import './chunk-CC2N4JGQ.js';
 import './chunk-OW7VVW6X.js';
 import './chunk-VBATFCWR.js';
 import './chunk-SG6ZTIMT.js';

package/dist/src/cli/main.js

@@ -1,10 +1,10 @@
 #!/usr/bin/env node
-import { DEFAULT_CONTROL_PLANE_URL, DEFAULT_DASHBOARD_URL, clearLocalCredentials, friendlyHostedError, runSessionCreate, runSessionList, runSessionStop, resolveCredentials, runTaskHosted, discoverRunSet, VERDICT_ARTIFACT_VERSION, persistCredentialsAfterLogin, DEFAULT_DOCS_SITE_ORIGIN, resolveSeams, resolveCachedAgentId, readLinkCache, postAgentResolver, writeLinkCache, ensurePomeGitignored } from '../../chunk-DINCNFAQ.js';
+import { DEFAULT_CONTROL_PLANE_URL, DEFAULT_DASHBOARD_URL, clearLocalCredentials, friendlyHostedError, runSessionCreate, runSessionList, runSessionStop, resolveCredentials, runTaskHosted, discoverRunSet, VERDICT_ARTIFACT_VERSION, persistCredentialsAfterLogin, DEFAULT_DOCS_SITE_ORIGIN, resolveSeams, resolveCachedAgentId, readLinkCache, postAgentResolver, writeLinkCache, ensurePomeGitignored } from '../../chunk-4LODFVKV.js';
 import { readManifest, writeManifest, MANIFEST_JSON, readRequiredManifest, normalizeManifestTwins } from '../../chunk-ABM3CMQB.js';
-import { assetPath, runTask, resolvePackageRoot, DEMO_TASK_NAME, demoTaskPath } from '../../chunk-ZUQYEPY4.js';
+import { assetPath, runTask, resolvePackageRoot, DEMO_TASK_NAME, demoTaskPath } from '../../chunk-QIVE6VJE.js';
 import '../../chunk-XDU6TD4O.js';
 import '../../chunk-CBFKZZBR.js';
-import { parseTaskFile, scoreStatus, runScoreLine, readLatestRun, readMetaSummary, outcomeOf, readConfigTwins, scoreCountsSummary, markerFor, criterionMarkerLabel, twinSkipSuffix, readCodeCriteria, createHostedClient, toTwinHttpEvent, redactJsonl, scoreFromFinalizeResponse, parseGitHubSeedState, uploadRunBlobs, isPreSatisfied } from '../../chunk-VBLY7LYM.js';
+import { parseTaskFile, scoreStatus, runScoreLine, readLatestRun, readMetaSummary, outcomeOf, readConfigTwins, scoreCountsSummary, markerFor, criterionMarkerLabel, twinSkipSuffix, readCodeCriteria, createHostedClient, toTwinHttpEvent, redactJsonl, scoreFromFinalizeResponse, parseGitHubSeedState, uploadRunBlobs, isPreSatisfied } from '../../chunk-XKBL7JKK.js';
 import '../../chunk-NW7HGA2K.js';
 import { MOUNTED_TWINS, deriveAgentSlug, exitCodeFor, HostedUsageError, HostedOrchError, HostedAuthError, HostedQuotaError } from '../../chunk-X66JOOO7.js';
 import { TAPE_ASSERTABLE_TOOLS } from '../../chunk-FKZZWWYC.js';
@@ -15,8 +15,8 @@ import '../../chunk-NJ246QPJ.js';
 import { LINEAR_CHECKS } from '../../chunk-XBT6ZLBG.js';
 import '../../chunk-ZKID2HS3.js';
 import { oneOf, defineCheck, repoRef, VACUITY_SENTINEL_NUMBER, childStatePath, VACUITY_SENTINEL, statePath, templateSlots, renderCheck, checksDigest, checkPattern, checkNearMissPattern } from '../../chunk-JWJYNAWI.js';
-import '../../chunk-DG2VDEQE.js';
-import { TWIN_NAME_LIST, isTwinName, createGitHubSmokeApp } from '../../chunk-4VVQMRWP.js';
+import '../../chunk-FGNIWUEU.js';
+import { TWIN_NAME_LIST, isTwinName, createGitHubSmokeApp } from '../../chunk-CC2N4JGQ.js';
 import '../../chunk-OW7VVW6X.js';
 import { isLegacyEventRow, eventSchema } from '../../chunk-VBATFCWR.js';
 import { redactEvent, redactSecrets } from '../../chunk-SG6ZTIMT.js';
@@ -4541,7 +4541,7 @@ var DEFAULT_AGENT_COMMAND = `node ${DEFAULT_AGENT_FILE}`;
 var MANIFEST_SCHEMA_URL = "https://pome.sh/schemas/v1/pome.json";
 var MAX_UNREADABLE_PATHS_SHOWN = 5;
 function readPackageVersion() {
-  if ("0.23.26".length > 0) return "0.23.26";
+  if ("0.23.27".length > 0) return "0.23.27";
   try {
     const here = dirname(fileURLToPath(import.meta.url));
     const candidates = [
@@ -4996,7 +4996,7 @@ function createProgram() {
         return;
       }
       {
-        const { runDoctorChecks } = await import('../../checks-UINTOJ2L.js');
+        const { runDoctorChecks } = await import('../../checks-WHR4ALIJ.js');
         const { renderDoctorReport } = await import('../../render-ZQQ4UMNO.js');
         const doctorReport = await runDoctorChecks({ mode: useLocal ? "full" : "hosted" });
         if (!doctorReport.ok) {
@@ -5034,7 +5034,7 @@ function createProgram() {
               taskForRuns.config.runs
             );
             if (k > 1) {
-              const { runTrialGroup } = await import('../../runTrialGroup-CL4EJZD2.js');
+              const { runTrialGroup } = await import('../../runTrialGroup-LBNWOZPA.js');
               const fileForRerun = relative(process.cwd(), file);
               const rerunCommand = defaultTask ? options.trials !== void 0 ? `pome run -n ${k}` : "pome run" : `pome run ${fileForRerun && !fileForRerun.startsWith("..") ? fileForRerun : file} -n ${k}`;
               const groupResult = await runTrialGroup({
@@ -5128,7 +5128,7 @@ function createProgram() {
         process.exitCode = 5;
         return;
       }
-      const { runDemo } = await import('../../runDemo-5Z7PIYTZ.js');
+      const { runDemo } = await import('../../runDemo-HFISTHIE.js');
       const result = await runDemo({
         apiBase: opts.apiUrl.replace(/\/$/, ""),
         dashboardBase: process.env.POME_DASHBOARD_URL ?? DEFAULT_DASHBOARD_URL,
@@ -5146,7 +5146,7 @@ function createProgram() {
   program.command("doctor").description(
     "Check the agent\u2194twin wiring: pome.json (or pome.yaml) present + valid, the local twin boots + serves, requests routed to the twin (not a hardcoded production host), egress floor active. On failure prints one named cause (file:line where knowable) + one concrete fix and exits non-zero."
   ).action(async () => {
-    const { runDoctorChecks } = await import('../../checks-UINTOJ2L.js');
+    const { runDoctorChecks } = await import('../../checks-WHR4ALIJ.js');
     const { renderDoctorReport } = await import('../../render-ZQQ4UMNO.js');
     const report = await runDoctorChecks();
     for (const line of renderDoctorReport(report, { passNote: true })) console.error(line);
@@ -5322,7 +5322,7 @@ function createProgram() {
   ).description(
     "Start a standalone twin as a long-lived foreground server (Ctrl-C to stop)"
   ).action(async (name, options) => {
-    const { runTwinStartCommand } = await import('../../twinStart-MVO22G7C.js');
+    const { runTwinStartCommand } = await import('../../twinStart-HJX7TYZL.js');
     await runTwinStartCommand(name, options);
   });
   twin.command("reset").argument("[name]", "Twin name (default: github)", "github").description("Reset standalone twin state").action(async (name) => {

package/dist/{src-6JNLHNYE.js → src-HRW7OZBS.js}

@@ -3996,12 +3996,12 @@ var mcp_tools_list_meta_default = {
   transcription: {
     readFrom: "twin-stripe's own tools/list, booted on defaultSeed() and driven through Hono's app.request, read immediately before this table became fixture-driven (F-1325). Re-projected on 2026-08-12 for F-1484: every tool's inputSchema was rewritten from stripeToolInputSchema() of its zod schema \u2014 a no-op for 25 of the 26, and create_refund.reason's enum for the one. Names, descriptions and ordering are unchanged, and test/mcp-tool-fixture.test.ts boots the twin and asserts it serves exactly these bytes.",
     contentOrigin: "packages/twin-stripe/src/tools.ts \u2014 a hand-written TypeScript array of 26 tools. Each inputSchema is z.toJSONSchema() of that tool's zod schema, frozen here verbatim including the $schema key it emits.",
-    comparedToUpstream: "never, and not comparable today. @stripe/mcp is a stdio-to-HTTP proxy that declares no tools of its own and the served listing is a function of the caller's Restricted API Key, so F-1326 recorded stripe as not-captured (fixtures/mcp-tools-list/stripe.status.json) rather than inventing a deployment-invariant table."
+    comparedToUpstream: "compared, and credential-scoped. F-1326 recorded stripe as not-captured because @stripe/mcp is a stdio-to-HTTP proxy that declares no tools of its own and the served listing is a function of the caller's Restricted API Key; that stopped being the whole story on 2026-08-10, when a golden was captured live off https://mcp.stripe.com (fixtures/mcp-tools-list/stripe.meta.json \u2014 live-wire-oauth, 11 tools). F-1469 read this table against it: of the 36 union names exactly 1 is shared (create_refund), twin_only=25, upstream_only=10. One of 36 is the honest measurement rather than a failed comparison \u2014 the endpoint decides its listing from the caller's credential, so that golden declares completeness credential-scoped and unguarded in BOTH directions. So a name this twin serves that the golden lacks is never evidence the vendor lacks it: read the 25 as not-compared, never as twin-only divergence and never as full coverage. The one compared name is load-bearing \u2014 F-1484 narrowed create_refund.reason to the enum that golden carries."
   },
   notes: [
     "Seeded from the code table rather than from upstream on purpose. F-1325 is name-neutral by construction: every byte this twin serves is identical before and after it."
   ],
-  canonicalFileSha256: "6a294341065eb90138f26654ad00bf7b0b3750ee6c2f12df82a38fb91df7854f",
+  canonicalFileSha256: "49c8d40fc29e99be704d81362e3c7b87d72cac58c35686ed3aae8dc166db3fd0",
   files: {
     raw: "mcp-tools-list.raw.json",
     canonical: "mcp-tools-list.canonical.json"

package/dist/twinHarness-H73PEFYS.js

@@ -0,0 +1,5 @@
+export { UnsupportedTwinError, bootTwin } from './chunk-FGNIWUEU.js';
+export { STRIPE_LOCAL_ACCOUNT_ID } from './chunk-CC2N4JGQ.js';
+import './chunk-OW7VVW6X.js';
+import './chunk-VBATFCWR.js';
+import './chunk-SG6ZTIMT.js';

package/dist/{twinStart-MVO22G7C.js → twinStart-HJX7TYZL.js}

@@ -1,5 +1,5 @@
-import { bootTwin } from './chunk-DG2VDEQE.js';
-import { isTwinName, TWIN_NAMES, defaultPortFor, TWIN_REGISTRY } from './chunk-4VVQMRWP.js';
+import { bootTwin } from './chunk-FGNIWUEU.js';
+import { isTwinName, TWIN_NAMES, defaultPortFor, TWIN_REGISTRY } from './chunk-CC2N4JGQ.js';
 import './chunk-OW7VVW6X.js';
 import './chunk-VBATFCWR.js';
 import './chunk-SG6ZTIMT.js';

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@pome-sh/cli",
-  "version": "0.23.26",
+  "version": "0.23.27",
   "description": "Digital-twin testing for AI agents \u2014 run tasks against resettable local or hosted twins and record tool-call traces for evaluation on pome.sh.",
   "keywords": [
     "ai",

package/dist/twinHarness-WVKOQ7UP.js

@@ -1,5 +0,0 @@
-export { UnsupportedTwinError, bootTwin } from './chunk-DG2VDEQE.js';
-export { STRIPE_LOCAL_ACCOUNT_ID } from './chunk-4VVQMRWP.js';
-import './chunk-OW7VVW6X.js';
-import './chunk-VBATFCWR.js';
-import './chunk-SG6ZTIMT.js';